Professional Documents
Culture Documents
net/publication/3296813
CITATIONS READS
97 693
1 author:
John Yen
Pennsylvania State University
340 PUBLICATIONS 7,322 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by John Yen on 25 February 2015.
ABSTRACT
Traditionaly, fuzzy logic has been viewed in the AI community as an approach for managing uncertainty. In the
1990's, however, fuzzy logic has emerged as a paradigm for approximating a functional mapping. This complemen-
tary mordern view about the technology offers new insights about the foundation of fuzzy logic as well as new chal-
lenges regarding the identification of fuzzy models. In this paper, we will first review some of the major milestones in
the history of developing fuzzy logic technology. After a short summary of major concepts in fuzzy logic, we dis-
cuss a mordern view about the foundation of two types of fuzzy rules. Finally, we review some of the research in
addressing various challenges regarding automated identification of fuzzy rule-based models.
1 INTRODUCTION
Fuzzy logic (FL) has been a somewhat controversial technology since its birth. However, the large number of
successful industrial fuzzy logic applications in 1990's, especially those developed in Japan, has generated an increas-
ing interest in FL. Fuzzy logic and artificial intelligence (AI) have at least one common objective: to develop compu-
tational methods that can perform reasoning and problem solving tasks that require human intelligence. However,
fuzzy logic has an additional objective: to explore an effective tradeoff between precision and the cost in developing
an approximate model of a complex system or function. While the issue of the cost was not considered an important
issue in early AI work, it has become important in the past decade due to an increasing interest in resource-con-
strained intelligent agents.
Cost
Utility
imprecise precise
Traditionally, fuzzy logic has been viewed as a theory for dealing with uncertainty about complex systems. A
modern complementary perspective is, however, to view fuzzy logic as an approximation theory. This perspective on
fuzzy logic brings to the surface the underpinning of the theory described above - the cost-precision trade-off. Indeed,
providing a cost-effective solution to a wide range of real world problems is the primary reason that fuzzy logic has
found so many successful applications in industry to date. Understanding this driving force of the success of fuzzy
logic will prevent us from falling into the trap of debating “whether fuzzy logic can accomplish what X can not
accomplish” where X is an alternative technology such as probability theory, control theory, etc. Such a debate is usu-
ally not fruitful because it ignores one important issue -- cost. A better question to ask is “What is the difference
between the cost of a fuzzy logic approach and the cost of an approach based on X to accomplish a certain task?”
Fuzzy Logic research in Japan started with two small university research groups established in late 1970s: one
was lead by T. Terano and H. Shibata in Tokyo, and the other lead by K. Tanaka and K. Asai in Kanasai. Like fuzzy
logic researchers in the U.S., these researchers encountered an “anti-fuzzy” atmosphere in Japan during those early
days. However, their persistence and hard work would prove to be worthwhile a decade later. These Japanese
researchers, their students, and the students of their students would make many important contributions to the theory
as well as to the applications of fuzzy logic [11].
In the 1980’s, several important industrial applications of fuzzy logic was launched successfully in Japan. After
eight years of persistent research, development, and deployment efforts, Yasunobu and his colleagues at Hitachi put a
fuzzy logic-based automatic train operation control system into operation in Sendai city’s subway system in 1987
[29]. Another early successful industrial application of fuzzy logic is a water-treatment system developed by Fuji
Electric. These and other applications motivated many Japanese engineers to investigate a wide range of novel
fuzzy logic applications. This lead to the fuzzy boom.
The fuzzy boom in Japan was a result of close collaboration and technology transfer between universities and
industries. Two large-scale national research projects were established by two Japanese government agencies in
1987: the better known of the two is the Laboratory for International Fuzzy Engineering Research (LIFE). In late
January 1990, Matsushita Electric Industrial Co. named their newly developed fuzzy controlled automatic washing
machine “Asai-go (beloved wife) Day Fuzzy” and launched a major commercial campaign for the “fuzzy” product.
This campaign turns out to be a successfully marketing effort not only for the product, but also for the fuzzy logic
technology. A foreign word pronounced “fuzzy” was thus introduced to Japan with a new meaning -- intelligence.
Many other home electronics companies followed Panasonic’s approach and introduced fuzzy vacuum cleaners,
fuzzy rice cookers, fuzzy refrigerators, fuzzy camcorders (for stablizing the image under hand jittering), camera (for
smart auto-focus) and others. This resulted in a fuzzy vogue in Japan. As a result, the consumers in Japan all recog-
nized the Japanese word “fuzzy”, which won the gold prize for the new word in 1990 [11]. This fuzzy boom in Japan
triggered a broad and serious interest in this technology in Korea, Europe, and, to a lesser extent, in the United States,
where fuzzy logic was invented.
Fuzzy logic has also found its applications in the financial area. The first financial trading system using fuzzy
logic was Yamaichi Fuzzy Fund. It handles 65 industries and a majority of the stocks listed on Nikkei Dow and con-
sists of approximately 800 fuzzy rules. Rules are determined monthly by a group of experts and modified by senior
business analysts as necessary. The system was tested for two years, and its performance in terms of the return and
growth exceeds the Nikkei Average by over 20%. While in testing, the system recommended “sell” 18 days before the
Black Monday in 1987. The system went to commercial operations in 1988.
The first special-purpose VLSI chip for performing fuzzy logic inferences was developed by M. Togai and H.
Watanabe in 1986 [27]. These special-purpose VLSI chips can enhance the performance of fuzzy rule-based systems
for real-time applications. Several companies were formed (e.g., Togai Infralogic, APTRONIX, INFORM) were
formed to commercialize hardware and software tools for developing fuzzy systems. Vendors of conventional con-
trol design software also started introducing add-on toolbox for designing fuzzy systems. The Fuzzy Logic Toolbox
for MATLAB, for instance, was introduced as an add-on component to MATLAB in 1994.
Another trend that contributed to research in fuzzy model identification is the increasing visibility of neural net-
work research in the late 1980’s. Because of certain similarities between neural networks and fuzzy logic, researchers
began to investigate ways to combine the two technologies. The most important outcome of this trend is the develop-
ment of various techniques for identifying the parameters in a fuzzy system using neural network learning techniques.
A system built this way is called a neuro-fuzzy system [15, 18].
The 1990’s is an era of new computational paradigms. In addition to fuzzy logic and neural networks, a third
non-conventional computational paradigm also became popular -- evolutionary computing, which includes genetic
algorithms, evolutionary strategies, and evolutionary programming. Genetic algorithms(GA) and evolutionary strate-
gies are optimization techniques that attempt to avoid being easily trapped in local minima by simultaneously explor-
ing multiple points in the search space and by generating new points based on the Darwinian theory of evolution --
survival of the fittest. The popularity of GA in the 1990’s inspired the use of GA for optimizing parameters in fuzzy
systems [13]. Various synergistic combinations of neural networks, genetic algorithms, and fuzzy logic help people to
view them as complementary. To distinguish these paradigms from the conventional methodologies based on precise
formulations, Zadeh introduced the term soft computing in early 1990’s [41].
A linguistic variable is a variable whose value can be described (1) qualitatively using an expression involving
linguistic terms and (2) quantitatively using a corresponding membership function [39]. The linguistic term is useful
for communicating concepts and knowledge with human beings; whereas membership function is useful for process-
ing numeric input data. A linguistic variable is like a composition of a symbolic variable in AI (a variable whose
value is a symbol) and a numeric variable (a variable whose value is a number) in science and engineering. Using the
notion of linguistic variable to combine these two kinds of variables into a uniform framework is, in fact, one of the
main reasons that fuzzy logic has been successful in offering intelligent approaches in engineering and many other
areas that deal with continuous problem domains. In general, the value of a linguistic variable can be a linguistic
expression involving a set of linguistic terms, modifiers such as “very”, “more or less” (called hedges) and connec-
tives (e.g., “and”, “or”). For example, the sentence “trading is moderate and not very heavy” can be expressed as
assigning the linguistic expression ‘‘Moderate AND NOT VERY Heavy’’ to the linguistic variable TradingQuantity.
The meaning of such an expression is governed by several semantic rules about constructing the corresponding mem-
bership functions.
When an interval is assigned to a variable with unknown value (e.g., the suspects age is between 20 and 30
years old), it constrains the possible values of the variable. Similarly, when a fuzzy set is assigned to a linguistic vari-
able (e.g., the suspects age is young), it imposes an elastic constrain on the possible values of the variable called the
possibility distribution [40]. The main difference between the two is that the notion of possible vs. impossible values
becomes a matter of degree.
One of the questions commonly raised about possibility distribution is its relationship with probability distribu-
tion. While the subject is too complex for a comprehensive discussion here, it is important enough to deserve some
clarification. The best way to understand the relationship between possibility distribution and probability distribution
is to compare interval-values with probability. As we have mentioned earlier, an interval-valued assignment con-
strains the possible value of a variable without indicating the likelihood that the variable takes a specific value in the
interval. Similarly, a possibility distribution states the degree of ease (i.e., possibility) for the variable to take a certain
value without indicating the likelihood that the variable has such a value. Even though possibility distribution and
probability distribution are different, they are also related -- if a value is impossible, it is obviously improbable. In
general, a possibility distribution can be viewed as an upper bound on the probability distribution. More importantly,
however, fuzzy logic can be viewed as complementary to the probability theory. For instance, the notion of events in
probability theory can be generalized to fuzzy events based on fuzzy sets. Such a generalization allows probability
theory to better deal with events that do not have a well-defined sharp boundary (e.g., the probability that the stock
Fuzzy set theory generalizes the conventional set theory; therefore, its axiomatic foundation is unavoidably dif-
ferent from that of classical set theory. More specifically, it has to violate two fundamental laws of Boolean algebra
-- the law of excluded middle A ∪ A = U and the law of contradiction A ∩ A = φ In other words, it is possible for
an element to partially belong to both a fuzzy set and the set’s complement. For instance, suppose that John is some-
what bald but not completely bald. In fuzzy set theory, John can partially belong to the set of bald people as well as
the set of people who are not bald. Due to the fact that the law of excluded middle and the law of contradiction are
not axioms of fuzzy set theory, formula equivalents in classical set theory are not necessarily equivalent in fuzzy set
theory. Similarly, logically equivalent formula are not necessarily equivalent in fuzzy logic. A potential danger of
ignoring such a difference is to reject fuzzy sets (and fuzzy logic in general) based on an inappropriate set of axioms
[8].
The set operations intersection and union correspond to logic operations, conjunction (and) and disjunction (or),
respectively. There are multiple choices for the fuzzy conjunction and the fuzzy disjunction operators. A common
choice is to use min for fuzzy conjunction, max for fuzzy disjunction; another common pair is the algebraic product
(for fuzzy conjunction) and algebraic sum (for fuzzy disjunction):
µ( A ∩ B) ( x ) = µ A ( x ) × µB ( x ) (EQ 1)
µ A ∪ B ( x ) = µ A ( x ) + µB ( x ) – µ A ( x ) × µB ( x ) (EQ 2)
The choice of a fuzzy conjunction operator could determine the choice of the fuzzy disjunction operator, and
vice versa. This is due to the principle of duality between the two operators. More specifically, a fuzzy conjunction
operator, denoted as t ( x, y ) and a fuzzy disjunction operator, denoted as s ( x, y ), form a dual pair if they satisfy the
following condition:
1 – t ( x, y ) = s ( 1 – x, 1 – y ) (EQ 3)
The set of candidate fuzzy conjunction operators, called triangular norms or t-norms, is defined by a set of axi-
oms. Similarly, the set of candidate fuzzy disjunction operators called triangular conorms, t-conorms, or s-norms is
defined by a set of dual axioms. We formally define t-norms and t-conorms using their axioms below.
DEFINITION 2 A t-conorm operator, denoted as s(x,y), is a function mapping from [0, 1] × [0, 1] to [0, 1] the
following conditions for any w, x, y, z ∈ [ 0, 1 ].
1.s ( 1, 1 ) = 1, s ( x, 0 ) = s ( 0, x ) = x
2. s(x,y) ≤ s(z,w) if x ≤ z and y ≤ w (monotonicity)
3. s ( x, y ) = s ( y, x ) (commutativity)
4. s ( x, s ( y, z ) ) = s ( s ( ( x, y ), z ) ) (associativity)
A summary of t-norm and t-conorm operators can be found in [14]. An important property about t-norms is that all t-
norms are bounded above by min and bounded below by drastic product. Similarly, all t-conorms are bounded above
by drastic sum and bounded below by max.
A fuzzy relation generalizes the classical notion of relation into a matter of degree. For instance, the fuzzy rela-
tion Friend could describe the degree of friendship between two persons. Similarly, a fuzzy relation Petite between
height and weight of a person describes the degree by which a person with a specific height and weight is considered
petite. Formally, fuzzy relation R between variables x and y, whose domains are X and Y, respectively, is defined by a
function that maps ordered pairs in X × Y to their degree in the relation, which is a number between 0 and 1, i.e.,
respectively, is defined by a function that maps an n-tuple < x 1, x 2, …, x n > in X 1 × X 2 × … × X n to a number in the
interval, i.e., R: X 1 × X 2 × … × X n → [ 0, 1 ] . Just as a classical relation can be viewed as a set, a fuzzy relation can be
viewed as a fuzzy subset. From this perspective, the mapping above is equivalent to the membership function of a
multidimensional fuzzy set.
If the possible values of x and y are discrete, we can express a fuzzy relation in a matrix form. For example, sup-
pose we wish to express a fuzzy relation Petite in terms of the height and the weight of a female. Suppose the range of
the height and the weight of interest to us are {5’, 5’1”, 5’2”, 5’,3”, 5’4”, 5’5”, 5’6”}, denoted h, and {90, 95, 100,
105, 110, 115, 120, 125} (in lb,) denoted w, respectively. We can express the fuzzy relation in a matrix form as shown
below:
Each entry in the matrix indicates the degree a female with the corresponding height (i.e., the row heading) and
weight (i.e., the column heading) is considered to be petite. Once we define the Petite fuzzy relation, we can answer
two kinds of questions:
• What is the degree that a female with a specific height and a specific weight is considered to be petite?
• What is the possibility that a petite person has a specific pair of height and weight measures?
In answering the first question, the fuzzy relation is equivalent to the membership function of a multidimen-
sional fuzzy set. In the second case, the fuzzy relation becomes a possibility distribution assigned to a petite person
whose actual height and weight are unknown. The second usage of fuzzy relation enables us to reason about the pos-
sible height of a petite person given her weight. For instance, we may wish to know the possible weight of a petite
female called Michelle who is about 5’4” tall where “about” indicates impression. The answer to this question can
be obtained through an important inference technique in fuzzy logic -- the compositional rule of inference. We will
use the example above to introduce the foundation of this inference technique.
How do we find out whether it is possible for Michelle to have a specific weight, say 110 lb? We need to con-
sider all possible heights of the person and see if a petite person with such a height can weigh 110 lb. In general, a
petitie person is possible to weigh w if and only if (1) the person is possible to have height h and (2) it is possible for
a person with height h and weight w to be petite. This can be expressed in logic as follows:
where Possible-weight (x, wj) is a predicate to test whether it is possible for person x to have weight wj, Possible-
height (x, hi), and Petite(hi, wj) are similar predicates in first order logic, ∧ denotes the conjunction operator,
∨ denotes the disjunction operator, and ↔ denotes bidirectional implication. As we mentioned earlier, both
possibilities and relations become matters of degree in fuzzy logic. Hence, we can replace the predicates in Equation
5 with possibility distributions Πheight(x)(hi) and Πpetite(hi,wj) a to obtain the following formula to infer the possibility
of a petite person’s weight, denoted Πweight(x) (wj):
where ⊗ and ⊕ denote fuzzy conjunction and fuzzy disjunction respectively. We define the compositional rule of
inference more formally below.
DEFINITION 3 Let X and Y be the universes of discourse for variables x and y, respectively, and x i and y j be
elements of X and Y. Let R be a fuzzy relation that maps X × Y to [0, 1] and the possibility
distribution of X is known to be Πx (xi). The compositional rule of inference infers the possibility
distribution of Y as follows:
ΠY (yj) = ⊕ ( Π X ( x i ) ⊗ Π R ( x i ,y j ) ) (EQ 7)
xi
The compositional rule of inference is not uniquely defined. By choosing different fuzzy conjunction and fuzzy
disjunction operators, we get different compositional rules of inference. We list two that are commonly used in prac-
tice:
There are two types of fuzzy rules: 1) fuzzy mapping rules, and 2) fuzzy implication rules. A fuzzy mapping rule
describes a functional mapping relationship between inputs and an output using linguistic terms, while a fuzzy impli-
cation rule describes a generalized logic implication relationship between two logic formula involving linguistic vari-
able and imprecise linguistic terms. The foundation of fuzzy mapping rule is fuzzy graph, while the foundation of
fuzzy implication rule is a generalization to two-valued logic. The inference of fuzzy mapping rules involves a set of
rules whose antecedent conditions form a fuzzy partition of the input space. We call such a collection of fuzzy map-
ping rules a fuzzy model. The inference of fuzzy implication rules are performed individually. Even though the infer-
ence results of these rules can be combined, the desired properties of their inference are described in terms of the
The distinction between fuzzy implication rules and fuzzy mapping rules has not been clear in the literature.
Until early 1990’s, fuzzy rules used in control systems have been viewed as a special kind of fuzzy implication rule
[17]. However, it is difficult to explain the use of conjunction operator in forming the “fuzzy implication relation” of
rules and the use of fuzzy disjunction in aggregating the conclusion of rules. This difficulty gradually lead to the crys-
tallization of the fundamental differences between the two types of rules. Zadeh, Kosko, Dubois and Prade and many
others contribute to this process through keynote speeches, books [16], conference and journal publications [6].
The most fundamental difference between the semantics of fuzzy mapping rules and fuzzy implication rules is
in their inference behavior. Even though these two types of rules behave the same when their antecedents are satisfied,
they behave differently when their antecedents are not satisfied. We will illustrate this using an example. Suppose x
and y are two integer variables taking values from the interval [0, 10]. Suppose we know that ‘‘if x is between 1 and 3,
then y is either 7 or 8’’. This sentence can describes at least two kinds of knowledge: (1) as a logic implication, and
(2) as an association. Assuming that we also know the value of x is 5, the logic implication will infer that y is
unknown (i.e., y can be any integer in the interval [0, 10]), but the association knowledge will not make any conclu-
sion regarding y. The two types of knowledge correspond to the two types of fuzzy rules: logic implication is the
basis of fuzzy implication rules, while association knowledge is the essence of fuzzy mapping rules. That is, fuzzy
implication rules generalize set-to-set implications; whereas fuzzy mapping rules generalize set-to-set associations.
The former was motivated to allow intelligent systems to draw plausable conclusions in a way similar to human rea-
soning; while the latter was motivated to approximate complex relationships (e.g., nonlinear functions) in a cost-
effective and easily-comprehensable way.
We will first use an example to illustrate the difference between set-to-set implications and set-to-set mappings.
This discussion will then set the stage for us to describe the difference between the theoretical foundation of these
two types of fuzzy rules. Suppose x and y are variables taking values from U = {a, b, c, d, e, f} and V = {r, s, t, u, v},
respectively. Suppose further that we know the following implication is true:
Such a set-to-set implication specifies a set of possible implications such as x = b → y = s , and a set of impossible
implications such as x = b → y = r . Notice that if the antecedent is known to be false, the implication is true
regardless of y’s value. Therefore, the following implications is possible: x = a → y = r . Hence, we can represent
the meaning of the set-to-set implication using the following matrix:
r s t u v
a 1 1 1 1 1
b 0 1 1 0 0
R ( x i ,y j ) = (EQ 11)
c 0 1 1 0 0
d 0 1 1 0 0
e 1 1 1 1 1
f 1 1 1 1 1
where an entry in the relation R(xi, yj) represents whether ( x = x i ) → ( y = y j ) is possible (“1” means possible, and
“0” means impossible). It is also easy to see that such a relation can be constructed by replacing x and y in Equation
10 with pairs of xi and yj and determine whether the resulting (i.e., instantiated) implication is true or false, i.e.,
1 if ( ( x i ∈ { b, c, d } ) → ( y j ∈ { s, t } ) )
R ( x i, y j ) = (EQ 12)
0 if ¬( ( x i ∈ { b, c, d } ) → ( y j ∈ { s, t } ) )
Such a set-to-set mapping specifies a set of possible association between the values of x and y. Such an association
can be naturally represented by a relation R ( x i, y j ) that describes whether input-output pair is a possible association
based on the mapping rule. Hence, the relation for the mapping rule can be represented as follows:
1 if ( ( x i ∈ { b, c, d } ) ∧ ( y j ∈ { s, t } ) )
R ( x i, y j ) = (EQ 14)
0 if ¬( ( x i ∈ { b, c, d } ) ∧ ( y j ∈ { s, t } ) )
It may be worthwhile now to point out the difference between the possibility relation of a mapping rule and that
of an implication rule. Fig 2 depicts such a difference for the rule “If x is A then y is B.” The points in shaded areas are
possible while those in the white area are not possible. The shaded area in Fig 2 corresponds to the “l”entries in the
implications relation in Equation 11.1 Fig 2 echoes a point we made earlier implication rules and mapping rules
1. Because the way a fuzzy implication is constructed, we need to turn it counter clockwise
90 degrees to see how Fig 2 (a) is its missor image.
FIGURE 2. A Pictoria l View of the Possibility Relation for (a) an Implication Rule and (b) a
Corresponding Mapping Rule
(a) (b)
y y
B B
A x A x
Having discussed the meaning of a set-to-set non-fuzzy implication, we can now consider an implication involv-
ing fuzzy sets (i.e., fuzzy implication):
rk ( x is A ) → ( y is B ) (EQ 15)
where A and B are fuzzy subsets of U and V, respectively. As in the previous example, this implication also specifies
the possibility of various point-to-point implications. The main difference here is that the possibilities are no longer
binary. Rather, they become a matter of degree. Therefore, the meaning of the fuzzy implication can be represented
by an fuzzy relation R defined as
R I ( x i, y j ) = Π I ( ( x = x i ) → ( y = y j ) ) (EQ 16)
where ΠI denotes the possibility distribution imposed by the implication. In fuzzy logic, this possibility distribution is
constructed from the truth values of the instantiated (i.e., grounded) implications obtained by replacing variables in
the implication (i.e., x and y) with pairs of their possible values (i.e., xi, yj):
Π r k ( ( x = x i ) → ( y = y j ) ) = t ( ( x i is A) → ( y j is B ) ) (EQ 17)
where t denotes the truth value of a proposition. It is easy to see that Equation (17) is a natural extension of Equation
(12).
The fuzzy relation of a fuzzy mapping rule in the form of Equation (15) represents the possibility degrees of
association between pairs of input and output values. Hence, it extends Equation (14) into the following one:
R r k ( x i, y j ) = Π r k ( ( x = x i ) ∧ ( y = y j ) ) (EQ 18)
The possibility distribution is thus determined from the membership functions of A and B:
value of the antecedent proposition “ x i is A” and the truth value of the consequent proposition “ y j is B.” For the con-
venience of our discussion, we will refer to these truth values as αi and βj, respectively, i.e., t (xi is A) = αi t (yj is B)
= βj. The truth value of the implication (xi is A → y j is B) is thus a function I of αi and βj:
Various definitions of implication functions have been developed. To compare and evaluate them, several intui-
tive criteria of desired inference results of fuzzy implications have been established. These criteria will thus form the
basis for evaluating and comparing different fuzzy implication functions. An example of such a criteriia is given
below:
Given: x is A → y is B
x is not A
Infer: y is V (unkown)
Notice that “y is V” (assigning the entire universe of discourse V to y) represents “y is unknown”. Other intuitive
criteria are summarized in Table 1..
TABLE 1. Intuitive Criteria for Reasoning Involving Fuzzy Implication x is A → y is B
The first family of fuzzy implication is obtained by generalizing material implications in two-valued logic to
fuzzy logic. A material implication p → q is defined as ¬ p ∨ q . Generalizing this to fuzzy logic gives us
t ( p → q ) = t ( ¬ p ∨ q ) . More specifically, fuzzy implications in this family can be generically defined as:
t ( x i is A → y j is B ) = t ( ¬( x i is A ) ∨ ( y j is B ) )
= ( ( 1 – µ A ( xi ) ) ⊕ µB ( y j ) ) (EQ 20)
t ( x i is A → y j is B ) = 1 ∧ ( 1 – ( µ A ( x i ) + ( µ B ( y j ) ) (EQ 21)
The second family of fuzzy implication is based on logic equivalence between implications p → q and
t ( x i is A → y j is B ) = t ( ¬( x i is A ) ∨ [ ( x i is A ) ∧ y j is B ] )
= ( 1 – µ A ( xi ) ) ⊕ ( µ A ( xi ) ⊗ µB ( y j ) ) (EQ 22)
An example of fuzzy implication in this family is Zadeh’s maximum fuzzy implication function:
t ( x i is A → y j is B ) = ( 1 – µ A ( x i ) ) ∨ ( µ A ( x i ) ∧ µ B ( y j ) ) (EQ 23)
which is obtained by using min for fuzzy conjunction and max for fuzzy disjunction.
The third family of fuzzy implication generalizes the “standard sequence” of many-valued logic and its variants.
The implication in these logic systems is defined to be true whenever the consequent is as true or truer than the ante-
cedent, i.e., t ( p → q ) = 1 whenever t ( p ) ≤ t ( q ) . This is an important property of many multivalued logic systems
because it allows the following tautology (i.e., a logic formula that is always true) in two-valued logic to be main-
tained in multi-valued logic: f → f where f is any formula. In other words, a logic formula always implies itself,
regardless of its truth value. The fuzzy implication function in this family can all be described in the following form:
t ( x i is A → y j is B ) = sup { α α ∈ [0,1] , α ⊗ t ( x i is A ) ≤ t ( y i is B ) }
µ A ( xi ) ≤ µB ( y j )
t ( x i is A → y j is B ) = 1 (EQ 25)
0 µ A ( xi ) > µB ( y j )
1 µ A ( xi ) ≤ µB ( y j )
t ( x i is A → y j is B ) = (EQ 26)
B( y j)
µ µ A ( xi ) > µB ( y j )
1 µ A ( xi ) ≤ µB ( y j )
t ( x i is A → y j is B ) = µ B ( y j ) (EQ 27)
---------------- µ A ( xi ) > µB ( y j )
µ A ( xi )
These three fuzzy implication functions came from, respectively, the standard sequence many-valued logic sys-
tem (denoted S n in the literature), a many-valued logic system proposed by Kurt Godel (denoted G n in the literature),
and a many-valued logic system J. A. Goguen introduced in 1969. Fig 3 shows graphically the function surface of the
five fuzzy implication functions we discussed.
Even though implication functions in multivalued logic systems can be used for constructing fuzzy implication
relations, approximate reasoning in fuzzy logic is fundamentally different from logic inference in multi-valued logic
approximate reasoning infers possible values of a variable, whereas multivalued logic infers the truth values of
propositions. The connection between the two was established by Equation 17. Even if we choose to use a fuzzy
implication function originated in a multivalued logic system (e.g., standard sequence, Godelian implication, or
Goguen’s implication), approximate reasoning still benefits from other important concepts and techniques in fuzzy
logic such as the compositional rule of inference, fuzzy relations, and possibility distributions. Without them, approx-
imate reasoning would not have been possible.
(a) Zadeh’s arithmetic fuzzy implication (b) Zadeh’s maxmin fuzzy implication
A -> B A -> B
1 1
0.5 0.5
0 0
1 1
0 0.5 0 0.5
B B
0.5 0.5
A 0 A 0
1 1
(c) Standard sequence fuzzy implication (d) Godelion sequence fuzzy implication
A -> B A -> B
1 1
0.5 0.5
0 0
1 1
0 0.5 0 0.5
B B
0.5 0.5
A 0 A 0
1 1
A -> B
1
0.5
0
1
0 0.5
B
0.5
A 0
1
Table 2 summaries how the criteria introduced in Table 1 are satisfied by the five fuzzy implication functions
based on sup-min composition (except that the sup-product composition is applied to Goguen’s fuzzy implication).
We applied the sup-product composition to Goguen’s fuzzy implication because the conjunction operator in Goguen’s
EXAMPLE 1 Let U and V be two universes representing numeric ratings (from 1 to 10) of redness and ripeness
of tomatoes, respectively. We denote the variable of these two ratings as x and y, respectively. Let Red be a fuzzy
subset of U defined as
We denote fuzzy implication relations obtained from the standard sequence and Goguen’s implication as Rs, and Rgg,
respectively. Suppose we know that a specific tomato is very red, we can express this information as ‘‘tomato is
VERY Red’’ where VERY is a hedge that modifies the meaning of a fuzzy set by taking the square of its membership
2
function: (i.e., µ VERY A = ( µ A ) ). Applying the sup-min composition to the standard sequence implication, we
obtain the following possibility distribution about the ripeness of the tomato:
The example above illustrates that standard sequence implication can satisfy Criterion II-1, while Goguen’s
implication can satisfy Criterion II-2. However, we should point out that these results require some conditions about
the membership functions in the implication rules. Even though these conditions are often satisfied in practice, we
should not overlook them. S. Fukami, M. Mizumoto, and K. Tanaka were the first to formally analyze the relation-
ship between various fuzzy implication functions and the intuitive criteria of fuzzy implication rules [9].
y
f (crisp function)
large f* (fuzzy graph)
small
1 0 x
1
small medium large
A fuzzy graph f * from X to Y is a union of Cartesian products involving linguistic input-output associations (i.e.,
pairs of “x is Ai” and “y is Bi”). Let f ∗ be a fuzzy graph described by a set of fuzzy mapping rules in the form of ‘‘IF
x is Ai then y is Bi.’’ The fuzzy graph can be expressed mathematically as:
f ∗= ∪A × B
i
i i (EQ 30)
In f*, + denotes the fuzzy disjunction. The Cartesian product of A and B, denoted by A × B , is defined as
An expression of the form A × B where A and B are words (fuzzy sets) is referred as a Cartesian granule [42]. Figure
depicts a fuzzy graph consisting of three fuzzy mapping rules:
The inference (i.e. interpolative reasoning) of such a set of fuzzy mapping rulesl is also based on compositional
rule of inference introduced earlier. Given an input ‘‘x is A’’ to the model, the inferred output of the model is a possi-
bility distribution B’ of y:
B′ = A′ o f ∗ = A′ ° ( ∪A × B )
i
i i (EQ 31)
where f ∗ represents the fuzzy graph of a given fuzzy model, ° denotes the compositional rule of inferenc.
where Ai, Bi, and Ci are fuzzy subsets of U, V, and W, respectively. The model’s fuzzy graph f* is expressed as
∪R = ∪(A ∩ B ) × C
n n
f ∗= i i i i (EQ 33)
i=1 i=1
where R i is the fuzzy relation of ith. rule, A i and B i denote cylindircal extension of A i and B j respectively (i.e., a
simple technique to extend these fuzzy sets into the space U × V ). Let A’ and B’ be fuzzy subsets of U and V respec-
tively. If we compose A’ and B’ with f* using sup-min composition and compute the union in Equation (33) using the
max operator, then composing inputs with f* (i.e., the entire fuzzy rule-based model) are equivalent to first compos-
ing inputs with individual rules in f* and then aggregating their composition results.
∪ ( A' ∩ B' ) ° R
n
( A' ∩ B' ) ° f ∗ = i (EQ 34)
i=1
We now state a theorem that shows how the Mamdani model can be derived from fuzzy graph and Equation 31.
THEOREM 2 (Mamdani Model) Suppose a fuzzy rule-based model maps X × Y to Z using a set of n rules in
the form of
and receives inputs in the form of x is A’ and y is B’ where A’ and B’ are fuzzy subsets of U and V. Suppose the fuzzy
inference of the model is based on a sup-min composition between inputs and a fuzzy graph that is defined using max
and min for all fuzzy disjunctions and fuzzy conjunctions operations. Then, the output of the fuzzy model (before
defuzzification), denoted by C’, is characterized by the following membership function:
n
µ C' ( z ) = max ( α i ∧ µ C'i ( z ) ) (EQ 35)
i= 1
where α i = sup ( µ A' ( x ) ∧ µ Ai ( x ) ) ∧ sup ( µ B' ( y ) ∧ µ Bi ( y ) ) and ∧ denotes the min operator.
x y
The foundation of the standard additive model is a fuzzy graph, the sup-product composition, and the use of
“addition” as a rule aggregation operator. We formally state this as a theorem below.
THEOREM 3 (Standard Additive Model) Suppose f ∗ is a fuzzy graph consisting of rules in the form of
If the inference of the model uses sup-product composition, “product” for all fuzzy conjunction, and “addition” for
rule aggregation, and centroid defuzzification, then the model’s output for crisp inputs x = x0; y = y0 is
n
z = Centroid
i = 1
∑
µ Ai ( x 0 ) × µ Bi ( y 0 ) × C i
(EQ 36)
The grid partition is the most commonly used fuzzy partitioning methods in practice (particularly in control
applications). The rule in Equation 32, for instance, partitions the input space this way. Wang and Mendel has used
this type of fuzzy partition in their procedure for fuzzy rule extraction from numerical data [28]. Fig. 5(a) illustrates a
typical grid partition in a two-dimensional input space. The grids shown in Fig. 5(a) are uniformly partitioned and
static, and the performance of the resultant model depends entirely on the initial definition of these grids. An adaptive
fuzzy grid partition can be obtained if we introduce some learning procedure in constructing the partition. Two typi-
cal learning procedures used in practice are the gradient descent method suggested by Jang and genetic algorithms
suggested by Karr [12, 13]. Fig. 5(b) gives an example of an adaptive fuzzy grid partition in a two-dimensional input
space. Grid partition (both static and adaptive) is convenient to use, but it may encounter serious “rule explosion”
FIGURE 5. Various Methods for Partitioning the Input Space: (a) Grid Partition (Static); (b) Grid
Partition (Adaptive); (c) Tree Partition; (d) Scatter Partition
(a) (b)
(c) (d)
Tree partition is another method used in partitioning input space. Fig. 5(c) gives an example of a tree partition in
a two-dimensional input space. A tree partition results from a series of guillotine cuts. By a guillotine cut, we mean a
cut that is made entirely across the subspace to be partitioned; each of the regions so produced can then be subjected
to independent guillotine cutting. At the beginning of the ith iteration step, the input space is partitioned into i
regions. Then another guillotine cut is applied to one of the regions to further partition the entire space into (i+1)
regions. There are various strategies to decide which dimension to cut and where to cut it at each step. Some are based
merely on the distribution of training examples; others take the parameter identification methods into consideration
[25]. Tree partition relieves the problem of rule explosion to a great degree, but it is not easy to use in practice. A
number of heuristics are usually needed to find a proper tree structure, and there may be difficulties involved in
designing an optimal tree partition.
Evaluating fuzzy models purely based on their fitness to training data, like any other modeling paradigm, can
potentially lead to the overfitting problem. One way to address this issue is to use information theoretic criteria to
evaluate fuzzy models so that the complexity of these models is explicitly taken into account in evaluating their opti-
mality [32]. An important benefit of fuzzy model is that its rules are interpretable because they capture a local rela-
tionship between the model’s input and output. However, this virtue may be lost if a fuzzy model is evaluated only by
its global performance. A remedy to this problem is to combine local learning (i.e., learning individual rules) with
globall learning (i.e., learning the entire rule set) [33].
8 SUMMARY
In this paper, we have summarized major concepts and techniques in fuzzy logic. We have also presented a
modern perspective about two types of fuzzy rules: fuzzy implication rules and fuzzy mapping rules. The latter
have been widely used in fuzzy logic control and other industrial applications. This new perspective not only clarifies
the formal foundation of these rules, but also sheds lights on how to deal with various challenges in identifying and
learning fuzzy rule-based models for high dimensional problems.
Acknowledgments
We would like to thank Henri Prade and Didies Dubois for fruitful technical exchanges regarding fuzzy logic
and artificial intelligence. We also would like to thank Reza Langari and Liang Wang for many fruitful discussions.
This research was partially supported by National Science Foundation Young Investigator Award IRI 9257293.